As the artificial intelligence landscape continues to resemble a high-stakes, high-tech game of poker, OpenAI, not one to be left behind, is stepping up its game with enhanced text-generating capabilities and, get this, lower prices.
OpenAI, today, took the wraps off its new and improved versions of GPT-3.5-turbo and GPT-4, the latter being the shiny new kid on the block. These models are now flaunting a new trick called function calling, and it’s not just for show. Function calling, according to OpenAI, is akin to giving a detailed recipe to these models, and in turn, they whip up the code to bring that recipe to life.
Picture this: function calling can be harnessed to create nifty chatbots that, instead of just answering questions, go a step further by calling on external tools, converting everyday language into database queries and squeezing structured data from text. These models, OpenAI assures us, have undergone rigorous fine-tuning to not just identify when a function needs to be invoked, but also to respond in a structured format that aligns with the function’s requirements. In plain English, developers can now bank on these models to deliver structured data with more dependability.
But wait, there’s more. OpenAI is also rolling out a new variant of GPT-3.5-turbo that comes with a substantially enlarged context window. Now, I know what you’re thinking, “What the heck is a context window?” It’s a measure of the chunk of text the model evaluates before it conjures up any additional text. Imagine this as the model’s short-term memory, and we all know the problem with short-term memory – it can be notoriously fickle.
Models with puny context windows tend to suffer from a digital form of amnesia, often veering off track, sometimes in ways that can be quite troublesome. The new GPT-3.5-turbo, however, boasts a context length that is four times larger than its predecessor, although you’ll need to cough up twice the price for this perk. In terms of capacity, OpenAI states that it can handle about 20 pages of text in a single run, which, while impressive, still falls short of Anthropic’s flagship model’s ability to process hundreds of pages in one go. A supersized version of GPT-4 with a 32,000-token context window is also being tested, but it’s currently a limited-edition model.
On the sunnier side, OpenAI is slashing the prices for the original GPT-3.5-turbo – not the one with the expanded memory, mind you – by a quarter. Developers can now enjoy the benefits of the model at a cost of around 700 pages per dollar.
The price cut extends to text-embedding-ada-002 as well, which happens to be one of OpenAI’s more sought-after text embedding models. Now, for those scratching their heads, text embeddings are handy in gauging the similarity of text strings, a tool that’s quite useful in searches and recommendations. The cost of text-embedding-ada-002 has seen a whopping 75% reduction, a move made feasible by OpenAI’s improved system efficiency. This efficiency is, unquestionably, a top priority for the firm, given the hundreds of millions it pours into R&D and infrastructure.
Shifting gears slightly, OpenAI has indicated that it’s focusing on refining and upgrading existing models rather than building new models from the ground up, post the release of GPT-4 in March. Sam Altman, OpenAI’s CEO, confirmed at a recent Economic Times conference that the firm hasn’t started on the successor to GPT-4yet, pointing out that there is still “plenty of work to do” before they can set the ball rolling on that project
